Harmony better than uniformity: A new pre-training anomaly detection method with tensor domain adaptation for early fault evaluation

Abstract

Aiming to identify incipient fault as earlier as possible, early fault detection (EFD) becomes the key step to trigger predictive maintenance. To lower deployment burden, pre-training techniques are employed to provide well-initialized features before fine-tuning different EFD tasks. Current large-scale pre-training models focus on seeking uniformity features to align every task’s characteristic. But in practical engineering, degradation data is usually small-scale and under irregular noise interference. The pre-trained uniformity features are easily biased and less discriminative to early fault. To address this concern, this paper proposes a pre-training anomaly detection method with tensor domain adaptation. Running on the classical contrastive learning architecture, the proposed method first builds hypersphere-formed representation of anomaly detection rule. With the merit of tensor decomposition in extracting the intrinsic information from raw signals, the proposed method second designs a new tensorized rule adaptation mechanism to learn the task-invariant detection rule from normal state data that are used for pre-training. The pre-trained feature representation is then obtained from the feature extractor in contrastive learning via back-propagation. This paper takes rolling bearing as the validation object. Experimental results on the IEEE PHM Challenge 2012 and XJTU-SY datasets show that the proposed method can effectively improve the convergence speed and detection accuracy on different EFD tasks with shallow model, deep model and transfer learning, and also prove that the harmony on task-invariant detection rule and task-specific information facilitates the universal applicability of EFD pre-trained features in small scale. Harmony is then believed better than uniformity for the EFD pre-training.